The CI caching playbook

Dependency caching is the fastest win available. See the dedicated dependency caching recipes page for per-ecosystem YAML for npm, pnpm, yarn, pip, poetry, maven, gradle, composer, cargo, go modules, and bundler. Most teams recover 30-60 minutes per day from this one change.

The actions/cache@v4 action provides a key-value cache store scoped to your repository. The most important detail: the 10 GB repository-wide cap. When the cap is reached, GitHub evicts the least recently used caches first.

- uses: actions/cache@v4
  with:
    path: ~/.npm
    key: ${{ runner.os }}-node-${{ hashFiles('**/package-lock.json') }}
    restore-keys: |
      ${{ runner.os }}-node-

Cache key strategy: hash the lockfile (package-lock.json, yarn.lock, pnpm-lock.yaml), not the package.json. Include the OS in the key. This gives you OS-specific caches that survive lockfile changes gracefully via the restore-keys fallback.

Docker layer caching is one of the highest-ROI optimisations for teams that build container images in CI. See Docker build optimisation for the full recipe including cache-from: type=gha and registry cache backends.

Turborepo remote cache stores task outputs (build artefacts, test results) in a shared cache. On a cache hit, Turborepo skips the task entirely and replays the output. For a 10-package monorepo, remote cache typically saves 30-60% of CI minutes.

Vercel hosts Turborepo remote cache for free for small teams. The self-hosted option uses S3 or R2 with a simple adapter. See monorepo strategies for cost comparisons.

# With Vercel Remote Cache
- run: npx turbo build test --filter='...[origin/main]'
  env:
    TURBO_TOKEN: ${{ secrets.TURBO_TOKEN }}
    TURBO_TEAM: your-team-name

Nx Cloud provides remote cache plus distributed task execution. The free tier allows 500 CI pipeline hours per month. Pro starts at $300/month for unlimited CI hours. For large monorepos (50+ packages), the distributed execution feature can be as valuable as the cache itself.

Nx affected commands (nx affected:build) are the monorepo equivalent of the paths filter: run only what changed.

Bazel remote cache stores build artifacts keyed on a content-addressable hash of inputs. At scale, a warmed Bazel remote cache can reduce CI time by 70-90% on incremental builds. The trade-off is Bazel setup cost: typically 2-6 months of migration for a large codebase.

Dropbox's published Bazel migration showed 10x build speed improvement on large C++/Python binaries and ~50% reduction in CI compute. See case studies for the full analysis.